From-scratch PyTorch implementation of DreamerV4 (Hafner et al., 2024): masked-autoencoder tokenizer, block-causal flow-matching dynamics with bootstrap curriculum, agent-token finetuning, and PMPO imagination RL. Hardened for TPU v4 / torch_xla with fixed-shape graphs, on-device RNG, and bounded compile-cache footprint.

🎥 Generate high-quality interactive videos in real-time using Causal Forcing, a method that elevates autoregressive diffusion distillation.